Sipcode

A hook intercepts bulky commands before Claude Code sees them and swaps in compact equivalents. On the project's own benchmark, git diff drops from 21,500 tokens to 1,260, and recursive grep falls by about 80%. The rewrite happens live, so you don't change how you work.

The metering layer reads your local Claude transcripts and reports where tokens went. sipcode why separates duplicate file reads, idle context, and output overhead, then totals what you could recover. That's where Claude Code token savings stop being a guess and start being a number, which is the whole point of running the tool.

Drift detection watches for the early signs of context rot. When your latest session behaves differently from your recent baseline, like tokens per turn jumping from 1,050 to 8,000, Sipcode says so and tells you how to fix it. It stays silent when nothing's wrong.

Sipcode makes no network calls during normal use. Your transcripts, code, and token data stay on your laptop, which matters if you work under a strict data policy. The tradeoff is that it can't notify you about updates, so you have to run sipcode update yourself.

Since v1.6.15, installing Sipcode mid-session works without a restart and without leaving stale context behind. That means you can add it to a session that's already running, which most competing approaches don't support.

The project publishes a 62.6% median token saving across a locked 20-task corpus, with a range from 37.4% to 80.6%. The corpus and the method behind it are public, so you can re-run the audit yourself rather than taking the number on faith.

Sipcode exposes 15 MCP tools that let an agent inspect its own token behavior. It's an unusual design choice for a context tool, and it points to a project aimed at people who'd rather query the data than read a static dashboard.